Sebastian Mallaby on OpenAI's Cash Crunch, the AI Bubble Debate, and the China AI Race

Sebastian Mallaby

Yeah, I do hold to my prediction. Back in January the burn rate was just crazy. Although OpenAI had good products and quite a lot of traction — 900 million consumers — they won't be able to charge money for most of that. Only about 5% of retail consumers were actually paying. If you look at a chart of where these users are, the US is the number two market, India is first, and the next three are Brazil, Indonesia and so on. These are not rich consumers — you can't charge them very much money.

So they had a business model that imagined they could throw money in all directions: a collaboration with Jony Ive to build a new form factor that would supplant the iPhone, Sora video-generation models, and so on — all very expensive. And the revenue side simply wasn't there. The burn seemed to me totally unsustainable. Even though Sam Altman is a magician when it comes to raising money, he wasn't going to raise the $660 billion that internal documents projected as the burn rate for the next five years.

Since then, there has been some good news. OpenAI recognised it had to get the burn rate down. It's pulled out of a bunch of data-centre building projects — Stargate and so on. It's cancelled Sora, the video model, which was a total money loser, and it's tried to impose some order on the chaotic management, though only half successfully. In the meantime, OpenAI is squeezed between Anthropic, which is much better at frontier enterprise applications like coding assistants, cybersecurity and agentic work, and Gemini from Google DeepMind, which has reached more retail consumers and monetises far better because Google has plugged AI into its search-advertising business, which is now doing more revenue than ever.

I think OpenAI is technically a good lab, but it's very hard to monetise when your product has a lot of competition, is kind of a commodity, and you're not terribly well managed. They've relied too much on the fake-it-till-you-make-it Silicon Valley tactic — smoke-and-mirrors fundraising. Look at the fundraising they announced earlier this year: the headline number was $122 billion, an astronomical figure. But when you dig in — and I'm amazed the press didn't point this out more — about two-thirds of that was promises conditional on a successful IPO, or payment in kind like compute access. The real money was a small share of the total. Why announce a $122 billion headline number when anyone who digs in can see it's rubbish? Because they're trying to head-fake investors into putting more money in, to persuade people they have momentum. They don't. The IPO delay is just the latest evidence that they took a big gamble but are behind where they say they are.

Sebastian Mallaby

100%. Everybody remembers the WeWork story — WeWork was this rocket ship back in 2019, it went out with a prospectus for its IPO, and people looked at it and said this is a joke, nobody wanted to buy the shares, and the IPO never happened. So you can fail by going for the IPO. OpenAI is in a very tough position: on the one hand it needs the IPO, because it can't raise enough money staying private; on the other, if it tries the IPO it may not succeed, and then it's really cooked.

Sebastian Mallaby

That's a good point, and I agree we don't know as much as we would if Anthropic were a public company with a public prospectus. What we do know is that Anthropic has always targeted enterprise customers — the type that actually pay for the product — and that it's been ahead on things like coding assistants and cybersecurity AI. Not that OpenAI is bad — it's not far behind — but I think Anthropic is the cutting edge on the applications enterprises are really willing to pay for.

Anthropic also hasn't been distracted into announcing a whole suite of retail-oriented initiatives that came to nothing. OpenAI announced it was going to do shopping, and that doesn't seem to have happened. It said it was going to do ads — not sure they've got terribly far. It generated Sora, which was a huge money loser. Anthropic never went down that path. So Anthropic has been way more laser-focused on the part of the market that makes sense — enterprise — and just better managed.

The other point is that Anthropic, among the frontier labs, is known for the lowest scientist churn. People go there, they believe in the mission, they believe in Dario Amodei as leader, and they tend to stay. They don't job-hop, whereas the other labs are subject to constant job-hopping, which is obviously disruptive.

Sebastian Mallaby

My view is that we have an OpenAI bubble but not a general AI bubble. For the reasons we've discussed, I think OpenAI is 50/50. It might work — I'm not saying I know they'll fail. I'm saying there's a 50% chance that by next summer we'll find they couldn't go public, couldn't raise enough in the private markets, and have to sell themselves at a discount to another company — Amazon, Microsoft, or someone else that wants an AI team, because technically OpenAI is a good team.

On the broader question: there's debate about whether enterprise customers are having an "oh my god" moment, realising these tokens are expensive. I've spent the last 18 months telling my own teams to go wild with AI, experiment, token max — the more tokens you use, the more you're showing you're AI-forward. Now it's, wow, this is expensive and I haven't seen a productivity gain yet, so what am I doing? There are lots of stories about companies imposing a middle layer between the user and the models, routing a simple query to a cheap, low-token model, and only a genuinely hard one to something expensive. So there's some sensible rationalisation happening about how customers spend on this technology.

But fundamentally, if you look at what's happened since ChatGPT's release, the clear story is unbelievably fast progress. When ChatGPT came out, it hallucinated non-stop. GPT-4 arrived six months later and cut about 80% of the hallucination. Then very long context windows, so you could put a whole Tolstoy novel into the model and query it. Then reasoning systems that could do maths and logic, which had been impossible before. Then agentic systems, coding assistants, cybersecurity systems. Now you've got bespoke AI autonomous scientists emerging. That is unbelievably fast progress. So I fundamentally think AI as a sector — and therefore the demand for semiconductors, data centres, all the things people worry about — isn't a bubble. It's for real. It's going to take some time for companies to figure out how to ration their people's token use sensibly, but they're going to keep consuming a lot of tokens.

Sebastian Mallaby

I don't agree that's the proper reading. The proper reading is that we're seeing a rationalisation of the market. Three or four months ago, across the whole US ecosystem, you had xAI trying to compete, Meta trying to compete, plus the big three — Google DeepMind, OpenAI and Anthropic. That's five, before you count Mistral in France, Cohere in Canada, and all the Chinese models. That's a lot of competition, and I don't think it consolidates down to a winner-takes-all, 2010s-social-media-platform outcome. But some consolidation is in order, so it ends up looking like cloud computing, with three or four big providers. Now we've got three leaders standing in the US, plus the foreign ones. That feels good to me for the sector's future business stability.

Sebastian Mallaby

We've seen plenty of examples of acquisitions, or more recently acqui-hires — a smallish AI company like Inflection, which Mustafa Suleyman was running, got sucked into Microsoft; Character.AI got sucked back into Google. There's a playbook here. OpenAI is a lot bigger than either of those, so it would be a more complex playbook, but the demand for AI talent and AI products, and the compute infrastructure underneath it, isn't going away — this is useful stuff people are going to figure out how to use productively.

So I don't know whether the whole of OpenAI gets bought by Amazon, Microsoft or some other acquirer, or there's some fancy acqui-hire where part of OpenAI gets sucked into a big company, or there's a splintering where the technical staff get individually hired into other labs. Who knows. What I'm saying is there's a fundamental problem with the way they're running their business model. I think they understand that — it's why they've pulled out of data-centre building and other things over the last six months — but they've got some way to go to fix and patch things up.

One of the lessons from how startups work, from my previous book The Power Law, is that when you have a very high valuation, a down round is super painful. They were valued in the last round at around $850 billion post-money, and in the secondary market they're trading for a lot less than that. If they were to just say, okay, we accept we're really worth $600 billion, the hit to everybody's equity options inside OpenAI would be horrible, and they would lose people. The hit to investors who'd believed in OpenAI would be bad, and they'd get pissed off, and the whole momentum machine that Altman has built would go through a convulsion. It might be what you have to do to make the thing sustainable, but my point is: once you ratchet all the way up to a very high valuation, it's difficult to climb down.

Sebastian Mallaby

First of all, I think your formulation — that they're cheating capitalism, doing an end-run around it — is a good, perceptive, and quite amusing insight, so thank you for that. I'd also say this is just the way the world is going, or at least the US. If you look at the number of American companies in which the US government has announced or consummated a deal — a colleague of mine, Jonathan Hillman at the Council on Foreign Relations, did a formal count, which just went up on the Council's website — the answer is 30 such companies since the Trump team came into power in January 2025, where the US government holds an equity stake in a private company.

So this is where the world is going, and the trend has been very much encouraged by the deceptive example of Intel. Look at Intel's performance since the government took a stake last August — it's been fantastic, way better than the Philadelphia Semiconductor Index, the normal comparable. Intel is up almost 400%; the index is up about 150%. Intel has done incredibly well since the government came in, and people lose sight of the fact that it did well because the Commerce Department is calling other companies and ordering them to do business with Intel. Intel gets a bunch of contracts and turns its game around because the government is picking a winner.

Now, it's one thing to say the government might have a justification for picking a winner when we have a problem with cutting-edge semiconductors all being made in Taiwan — we don't want to be reliant on an island that could be invaded by China, so we want domestic semiconductor manufacturing. I get that argument. I don't believe in extending it to OpenAI, which is just one of multiple American foundation-model builders. We don't need OpenAI for any strategic reason, so there's no justification for picking a winner around it. Capitalism is sometimes justifiably twisted because of a national-security reason. Backing OpenAI would not be a justifiable instance.

Sebastian Mallaby

I'd say welcome to China. That's the kind of thing the Chinese government would do — prop up the stock market with intervention of that sort. In the United States, when the Federal Reserve operates a policy that even looks like it's about stabilising the stock market, people freak out and call it a Fed put, and say it creates more bubbles down the line — capitalism doesn't work unless there's real risk involved. If you have a bunch of political types in Washington — the Commerce Department and so on — picking winners and distorting market outcomes, you don't have a market anymore. It's not free. To put your point differently: this is an end-run against the idea of a fair, level playing field on which companies compete and the most efficient ones win. That's supposed to be the wellspring of efficiency in American capitalism. If you start tilting the playing field by picking OpenAI as a winner, you've just trashed that.

Sebastian Mallaby

We haven't mentioned this yet, but I'm going to mention it now, since you've given me the excuse — I published a book this year called The Infinity Machine, about Demis Hassabis and—

Sebastian Mallaby

—DeepMind. There you are.

Sebastian Mallaby

Go read The Infinity Machine, folks. But no, seriously — the thing about China is it does everything faster. They got my manuscript last, then had to translate it into Chinese, then wanted photographs and other embellishments, so they produced a more complex product, but they actually published it before Penguin Press in the United States, or any of my other deals elsewhere.

So I go to China right at the start of my book tour. I spent eight days across four cities — Hangzhou, Shenzhen, Shanghai, Beijing — talking to computer scientists at private labs like Huawei and Ant Group, and to academic computer scientists at universities. What struck me first is that they talk about safety — they bring it up unprompted. The notion I've heard from friends in Washington, that the Chinese don't give a damn about AI safety, is just not true. They do talk about it. I'm not claiming government policy is to pursue safety, or that the majority view in China wants safety — China is like the US, it has accelerationists and people who want to go slower because they're worried about safety. Neither side is going to de-escalate unless the other does too. But it's wrong to caricature China as 100% acceleration. There is scope to talk to them about safety.

The other thing I observed is that China is very good, and very focused, on applications. Go to a company like Hikvision, which is under US sanctions, and it's an out-of-body, double-take experience — on one hand it feels like an American tech company. I love tech companies, all about building cool things and making the world better — I drink that Kool-Aid, I believe in it. I see these people building cool technology. They showed me an AI scanning camera you point at water to get a reading on pollution levels — and because they can measure it, there's now an internal market in water-pollution reduction between different Chinese cities. The downstream city pays the upstream city to reduce the pollution flowing down to it. You can only do pollution reduction once you can measure the pollution, and that's what this company is building.

But they're also under sanctions, because the US says — and historically this was true — that they build other kinds of cameras good for surveillance of civilians in Xinjiang and elsewhere. So they're both bad guys and cool guys — a difficult thing to reconcile. But whatever they are, they aren't going away. These are real companies building real technology. Go to Huawei and they've got application after application — a special AI to service the bullet train between Shanghai and Beijing every evening. We used to have human mechanics go under the train to check it; now they just have AI cameras and a couple of robots that fix it. They're doing this. We're not stopping them. We've imposed chip export controls on China to hold them back, and it hasn't worked — they're moving ahead.

The latest thing is a model from a group called JIPU in China, which isn't quite as good as mythos from Anthropic, but it's pretty close. We're kidding ourselves if we assume away the reality of China as a technology superpower. We need to get our heads out of the sand and start talking to China about what happens when they have a mythos-level model that could hack every single bank in the global financial system and wreak havoc. We need to persuade them not to release it on an open-source, open-weight basis, because then any criminal can use it, and there won't be an off switch.

Sebastian Mallaby

I believe in American power, first of all. I work at the Council on Foreign Relations in New York, we do geopolitics all day long, and I believe US power is generally a force for good. So I'd rather the Chinese were behind on AI. To the extent a chip export ban helps us stay ahead, I support it — when it was first announced in 2022, I wrote a long essay in the Washington Post about why it was a good idea. But my doubts recently come from looking at the results: I'm not seeing Chinese models that far behind. And because they're not far behind, we have to reckon with the reality that they're building models that will destabilise the global cyber system. Unless we persuade them not to release these on an open-weight basis, which is what they're doing now, we have serious trouble on our hands — everything in cyberspace gets destabilised, and we need a policy to deal with that proliferation risk. I'm in favour of the chip export ban if we could have it for free, with no downside. But if the effect of chip export controls is that we can't talk to them about an agreement on not releasing open-weight mythos-level models, I'm willing to trade a bit on the export ban.

Sebastian Mallaby

It shows they're making good models that serious American companies are considering using. That's another argument for why you can't just pretend China can be beaten and leave it there — these guys are for real, and we have to work with them, not just against them.

It's useful to think through the lens of the Cold War. With nuclear weapons, there were two kinds of big risk: one was a nuclear conflagration between the Soviet Union and the United States, which we prevented through mutually assured destruction — close parity in the two arsenals, and therefore deterrence. The other was proliferation of these weapons to rogue states or terrorists, which we dealt with through a separate mechanism, the non-proliferation regime. We were competing fiercely with Russia — an arms race, the Cuban Missile Crisis, being told at the UN "we will bury you," as Khrushchev said while banging his shoe on the table. So there was deadly serious competition between the two superpowers, but also cooperation on non-proliferation. I see the AI future the same way — inevitable competition between China and the US, but also, I hope, collaboration, because the proliferation risk is too awful to contemplate without it.

Sebastian Mallaby

I don't think you are. I think the Chinese have shown a pretty strong track record of stealing intellectual property from the US and then monetising it on their own terms.

Distillation is a process where, once a strong new American model comes out, a Chinese copycat asks it a ton of questions and harvests the answers as training data — if the question is like this, the answer should be like that. When a first-mover American lab has to train a model in some complicated frontier expertise — say quantum physics — it expensively hires a bunch of quantum physicists to create problem sets, model questions and answers, generating that training data. That's a super expensive, time-consuming, painful process. But once you've built an AI that can replicate all those quantum physicists, the Chinese can come along, skip hiring the human physicists, and just query the machine equivalent. That's distillation.

When Chinese companies do this, it isn't illegal, but it is against contract — when you sign up to use an American model, you agree not to query it endlessly and train your own model by copying its outputs. So they're violating contract, not federal law, as I understand it. Whatever the legal niceties, the question is: can you stop it? I'm all in favour of stopping it if we can, and it seems to me Anthropic, Google and OpenAI all have every commercial incentive to build in anti-distillation safeguards if they can find some. So it's a self-solving problem, insofar as it has a solution. And I should add — Elon Musk casually admitted a few months ago that his company xAI had distilled from one of the US frontier competitors. So it's not just the Chinese who do this. It's the rough and tumble of the marketplace — not nice, but the practical question is: stop it if we can. Insofar as we can't, we have to live with the reality that Chinese models are good.

Sebastian Mallaby

I'm not sure. By the way, I think the latest Anthropic models do have some anti-distillation technology built in, so we'll see how effective that turns out to be — it's going to be a cat-and-mouse game, both sides trying to get smarter. But to answer your question: let's posit US labs figure out how to stop distillation. Would the Chinese fall behind a lot, or just a bit? I don't think anybody really knows. I suspect that if they needed to generate their own data, they would — they'd pay more, it would take more time, but they'd get there, because they've got plenty of extremely smart Chinese scientists they could put to work generating training data.

Sebastian Mallaby

Embedded in the story of Google DeepMind and Demis is a kind of morality tale about somebody who really wanted to make AI safe — that was his driving passion from the moment he founded DeepMind in 2010. He bonded with his co-founder Shane Legg at a safety lecture where they discussed the potential for AI to attack humanity by the year 2030, which turns out to be perhaps a prescient projection — at least the capability will be there, whether the AI attacks is a different question.

Demis Hassabis was thinking about safety from the beginning, and when he sold his company to Google in 2014, a condition of the sale was a safety and ethics oversight board — he said he couldn't let AI be rolled out into the world just on the say-so of the Google corporate board, that there had to be independent people from outside. They talked about Barack Obama as an example — when he was leaving the presidency, could someone of that stature sit on a board saying when it's safe to roll out? That was Demis's vision, coupled with another hope: that all the major scientists would come together in one single effort to roll AI out, so there'd be no competitive pressure to move unsafely and too fast.

It turns out, and this transpires through the story I tell, that all of Demis's optimistic ideas about making AI safe crashed and burned. The idea of just one lab building AI turned out to be a pipe dream — humanity is too tribal, competitive and disputatious. There will be multiple labs; when you're confronted with the prospect of building a god machine, there'll be plenty of different sects of worshippers trying to do it. And the idea of oversight within Google — ultimately the Google board would not agree to giving outside grandees a veto over how they used a technology they were spending billions of dollars developing. Given their fiduciary obligations, they weren't going to do that. They couldn't.

So the experiment Demis ran at DeepMind — and I discovered internal documents with the back-and-forth red lines between two teams of lawyers over the exact safety mechanisms they might use, and all the secret strategising Demis did, threatening to spin out of Google if he didn't get the safety oversight he wanted. Then the Google DeepMind general counsel threatened me and said I wasn't allowed to publish any of it, and I said the heck with you, I'm publishing it anyway. It was all quite dramatic, but the bottom line is: it turns out to be impossible to impose safety restraints within one AI lab when that lab is in a competitive posture with respect to others. We saw the same thing play out at OpenAI, more publicly, when the safety board temporarily fired Sam Altman for five days.

So what this shows is: if you want to stop a race with multiple players, you need government to enforce restraint on all the players at once. And if there are players in China, you need the Chinese government to buy in too, and agree to put restraints on their labs when the US restrains its own. France, Canada — that's fine, the US can compel compliance there, because Cohere in Canada or Mistral in France depend on American technology and the American market to function. With China, you can't compel them. So there need to be two governments doing a deal, where everybody agrees to some caution and checking of models before release. It was the policy of the US government to do none of that — they even issued an executive order banning states from regulating AI in their own way.

Sebastian Mallaby

Given my view that government needs to get involved, I've been cheered up by what's happened since April, when mythos first came on the scene and galvanised the US government into paying attention and restricting release. Although you could argue, correctly, that on paper the executive order is a voluntary collaboration system with the frontier labs, the reality is it's not voluntary in the least. Commerce recently called up Altman at OpenAI and ordered him to seek government permission before giving his latest model to any customer — government has to sign off customer by customer. That's extremely heavy-handed. So I think they're in it for real — the government has realised it can't let this disseminate around the world without government control, and we're going to get pretty tough controls.

I think the gap in the system is that they're not talking about doing this in coordination with China, because the US policy world has two kinds of China expert: the ones who are always hawkish, and the ones who used to be hopeful about collaborating with China but flipped to being uber-hawks once Xi Jinping rose to power and frustrated those hopes. So you've got traditional hawks and new hawks, but both hawkish, and nobody wants to say they want to talk to China. That's the huge gap in the posture. The US government has done a 180 on domestic regulation, and I welcome that. The next thing that has to come, just because it's necessary, is getting over the inhibition about talking to China.

Sebastian Mallaby

It's a bit more complicated than that. I think you can talk and put pressure on them at the same time. Going back to the Cold War analogy, there was vicious competition between the Soviet Union and the United States at the same time as collaboration over proliferation. There will be competition — and there are ideas around strengthening chip export controls that I'm not against. Economists sometimes talk about corner solutions: you can have a fully pegged currency or a fully floating one, but if you go for some mushy semi-pegged middle ground, hedge-fund speculators will see you're not really determined to defend it, and they'll eat you for breakfast, lunch and dinner. It's the same with AI policy — there are corner solutions. You could give up the export controls, or offer to loosen them as a show of sincerity, in return for collaboration on fixing the non-proliferation risk. That's one corner solution. Or you don't say that — to the contrary, you tighten up the chip export controls. There's a massive loophole right now: a Chinese model builder can train on the most advanced Nvidia chips all day long, because the cloud compute they access is in Malaysia or some other offshore place that's fully allowed to import those chips. It's nuts that loophole exists — you tell the Chinese they can't use Nvidia chips, then let them use a data centre just across the border. So the other corner solution is to get serious, cut off the loophole, cut off the distillation, and put China in a position so weak it's begging for collaboration. I'm agnostic — I think we need to collaborate, and I'm flexible on how we get there.

Sebastian Mallaby

Simply because mythos was so powerful, it was very threatening. The prospect that you could take this model, find code vulnerabilities in every single entity on the internet, and then hack them — that's curtains for the financial system. That's why they took it seriously.

Sebastian Mallaby

Well, you've got this sort of Putin-and-the-oligarchs story — endless examples of authoritarian governments and big business titans, and where does the power really lie, how stable is that relationship? The answer is it tends not to be stable, point one. And point two: the government wins, because it has the monopoly on coercion. I think Silicon Valley is figuring that out, realising the government is too powerful to ignore. Dario Amodei tried to tell the government it shouldn't use these tools for certain things, like mass domestic surveillance, and the government said get lost — we'll call you a supply-chain risk, we're not going to be dictated to. Who won that fight? Clearly the government did.

Sebastian Mallaby

Absolutely.

Sebastian Mallaby

You've got Sam Altman, who is essentially a commercial opportunist — he wants to win commercially, or at least survive commercially. His drive is to be a big shot; he thought about running for governor of California at one point, being a political big shot, but decided building AI was a bigger big shot, and he wants to put his imprint on it. He's not a PhD scientist — he doesn't even have a first degree, having dropped out of Stanford. That's not to say he isn't massively smart, but he isn't a deep scientist.

Then you've got people like Dario Amodei and Demis Hassabis, PhD scientists who come at this wanting to use AI to advance deep science. That's their deepest motivation, and I believe it's very deep with both of them — I believe that's the reason they're number one and number two in this race. It's good for recruiting the best scientists, and good for holding together and leading a fundamentally scientific enterprise like building artificial general intelligence.

The point where this really came home to me was when I was talking to Demis one day about his motivation for building AI, and he said: "Listen, when I'm reading scientific papers at two in the morning, Sebastian, I see reality staring me in the face, calling to me, saying, 'I'm here to be discovered.' And if I had artificial general intelligence, I could discover the fundamental rules that explain the fabric of reality. It would be like understanding all of nature, which presumably may have been created by some kind of divine intelligence. So in this sense, my quest for AGI is kind of like my way of getting closer to what I might call God."

Sebastian Mallaby

Thank you so much. Nice to talk to you.